Echo

2024-present

Download on the App Store

The Story

As the state-of-the-art models are released, I find myself so impressed by their raw capabilities. But as much as I want an LLM that can code as well as I can, what I really want is a model that codes, thinks, and writes like me. I want a KevinGPT.

Seeing your training progress; having a conversation with the Echo interviewer; reading your custom news summary

Echo is my attempt to create just that. It uses ongoing voice conversations with the Echo bot to gather training data to inform the model how to think more like you. To start, you answer 20 general questions like “What is a dream of yours?” After that, customized questions just for you are generated to keep the conversation going and to keep your echo learning about how you think, process, and respond to new data.

After your own echo is fine-tuned to your personality, you can send podcasts and articles to it to summarize, and even filter that summary to either confirm your values or open your mind. You can also export your trained echo as a system prompt to use with ChatGPT or Claude.

2-minute demo of Echo

What I’ve Learned So Far

I’ve learned a few new technical skills with this project: fine-tuning with LoRA, transformer architecture, audio/transcript transport to minimize latency.

I’ve also learned that this is an incredibly difficult problem to solve by myself. It’s practically a full time job implementing the latest models that seem to be released every 3 days at this point.

But, just because it’s hard, doesn’t mean I won’t keep chipping away at it…

My Role

• Design
• SwiftUI
• Ruby on Rails
• LLMs

I designed and built Echo by myself, with product advice from close mentors and investors, like Tim Kendall (from my Moment days) and Evan Tato. It’s a SwiftUI native iOS app with a Ruby on Rails backend. I’m fine-tuning various open weight LLMs to train your echo.

What I’m Most Proud Of

How these conversations with a robot feel natural. Initially, you’re tasked with spending 100+ minutes talking to Echo to train it, and it’s actually an interesting and engaging experience. The conversation isn’t perfect, but it’s constantly improving and has come so far from when I first launched Echo in late 2024.

Gemini Live speech-to-speech is the keystone to make the conversations feel authentic, so I can’t take all of the credit here. I’ve had a lot of fun watching Echo grow with the raw capabilities of the new speech and thinking models.